See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330538122

Can Modern AI replace teachers? Not so fast! Artificial Intelligence and Adaptive
Learning: Personalized Education in the AI age

Article · December 2018

DOI: 10.21692/haps.2018.032

CITATIONS READS

27 5,509

1 author:

Vasiliy Kolchenko
New York City College of Technology
16 PUBLICATIONS 632 CITATIONS

SEE PROFILE

All content following this page was uploaded by Vasiliy Kolchenko on 06 June 2019.

The user has requested enhancement of the downloaded file.

Can Modern AI replace teachers? Not so fast! Artificial Intelligence and Adaptive
Learning: Personalized Education in the AI age

Vasiliy Kolchenko*
*Corresponding Author: vkolchenko@gmail.com
HAPS Educator. Vol 22 (3), pp. 249-252. Published December 2018.
doi: 10.21692/haps.2018.032

Kolchenko V. (2018). Can Modern AI replace teachers? Not so fast! Artificial Intelligence
and Adaptive Learning: Personalized Education in the AI age. HAPS Educator 22 (3):
249-252. doi: 10.21692/haps.2018.032
 Can Modern AI replace teachers? Not so fast!
Artificial Intelligence and Adaptive Learning: Personalized
Education in the AI age

Vasiliy Kolchenko, MD/PhD

Department of Biological Sciences, New York City College of Technology, The City University of New York, 300 Jay St., Brooklyn NY
11201
vkolchenko@gmail.com

Abstract
Artificial intelligence (AI) is transforming many fields, including education. Can modern AI replace teachers? We discuss some
popular AI applications for adaptive learning, how they are related to personalized education, what makes these applications
“intelligent,” and how the quality and quantity of student-generated data changes the behavior of adaptive learning systems and
the learning experiences of students. If the data are scarce and corrupt, the benefits of adaptive learning are minimal. Motivated
and persistent students generate greater amounts of high quality data and, as a result, tend to have better learning experiences.
This can increase the gap between high achieving and low achieving students. More realistic student models and better
understanding of the pedagogical context are needed to improve the performance of the AI educational systems. These are some
of the reasons why experienced teachers cannot be replaced by the current AI applications.
https://doi.org/10.21692/haps.2018.032

Key words: artificial, Intelligence, adaptive, learning, AI

Is AI everywhere?
We live in the age of artificial intelligence (AI). Pervasive,
but mostly hidden, it is taking over many jobs, devices,
technologies and media, promising to transform all fields, from
industry and commerce to medicine and education (Hawkins
and Blakeslee 2007, Nilsson 2009).
Increasingly, our actions and decisions are guided by AI
(Bostrom 2014). It tells us how to drive to our destination
(GPS), what to buy (online shopping and ads), what to read
(online search) and whom to date (dating sites).
Can it also tell educators how to do their jobs better, or
even replace teachers completely? After all, AI is about to
start replacing many professionals, from truck drivers to
radiologists.
Reliance on AI is inevitable because in many situations modern
AI is more effective and much more efficient than people.
Not only can it win in chess and Jeopardy, it can also evaluate
radiological images better than radiologists and avoid traffic
accidents better than truck drivers, saving lives in both cases.
In education, AI promises to deliver what has always been
the highest goal of pedagogy, wise and caring guidance for
each student, adapted to the individual’s needs. AI would
potentially provide customized learning resources and
activities, combined with the pace and style of instruction that
best suits each individual student (Luckin et al. 2016).
As a result, all learners would have their own highly intelligent
digital tutor (Andriessen and Sandberg 1999, Beck et al. 1996)
and the problem of personalized education could be solved
once and for all. The remaining teachers would supervise and
adjust this computational adaptive learning.
But can modern AI actually deliver this? What can AI do at
the moment and what are the problems that slow down its
progress?
What is AI?
AI is the computing that aspires to do what people do, only
better. It encompasses perception (computer vision and
hearing), reasoning (understanding and generating language
in forms such as Siri and Alexa), learning, prediction, making
decisions, and performing actions (robots and self-driving
cars).
To do all of this intelligently, modern AI needs effective
learning algorithms and a lot of high quality data for algorithm
training (Russell and Norvig 2010, Domingos 2015). The most
popular and successful algorithms today are artificial neural
networks, biologically inspired programs based on the neural
networks of the brain.
The data are specific to the applications. For example, all our
choices online – what we click, read, watch, and buy – are
valuable data for the AI algorithms that are trying to figure out
how to serve us better (and also how to sell us more goods
and services).

continued on next page

249 • HAPS Educator Journal of the Human Anatomy and Physiology Society Volume 22, Issue 3 December 2018
 Can Modern AI replace teachers? Not so fast! Artificial Intelligence and Adaptive Learning: Personalized Education in the AI age
The data feed learning and learning improves the program’s
intelligence, which means that the AI makes better predictions
and decisions. The more high-quality data are available for
training, the better the outcome. This continuous cycle of self-
improvement reinforces the program’s abilities. It is similar to
some positive feedback mechanisms in physiology.
But if the data are scarce or confusing, the AI program may
perform progressively worse, becoming more clueless instead
of smarter and smarter. For example, if you are clicking on the
online ads at random, you are providing AI with meaningless,
confusing data. In this case, AI needs to recognize this
immediately. Otherwise it will be learning from corrupt data,
which may lead to wrong predictions and bad decisions.
Artificial intelligence specialists informally call this “garbage in,
garbage out.”
Something similar can happen in adaptive learning. In fact,
it is quite likely to happen, considering what students are
supposed to do while using many popular programs.
What is adaptive learning?
Adaptive learning (AL) is an educational approach that uses
computing algorithms in order to organize the interactive
learning experience and adapt to the individual needs of
each student. Therefore, it is part of personalized education.
Another type of personalized education is, for example, one-
on-one tutoring, with no computers involved. This type of
personalized education has been around since the times of
Socrates, and probably even predates Socrates.
Modern AL is mostly intelligent. This means that it uses
AI. There are some straightforward computer algorithms
in educational programs that do not use AI, but they have
become rare. Not all adaptive learning systems are intelligent,
and not all intelligent learning systems are adaptive, but for
the most part they intersect (Brusilovsky and Peylo 2003).
A simple example would be a computer program that
provides various learning resources and engages students in
the interactive learning (Sleeman and Brown 1982). When a
student makes a mistake answering a question, the program
does not simply give him another similar question. Instead,
it searches for the patterns in all of the student’s prior
interactions with the program and analyzes them. Based
on this, it creates a model of the student and uses it for the
optimal response (Brusilovsky and Millan 2007).
This response, which might be a suggestion to review a
relevant paragraph, to watch a short video, or to answer a
series of increasingly difficult questions, depends not on just
one wrong answer but on the complete knowledge about the
student that the program has accumulated so far. It is different
for each student because each student has interacted with
the program in a different way. The program remembers all
250 • HAPS Educator Journal of the Human Anatomy and Physiology Society
of these interactions and makes its own conclusions about
the appropriate resources, pace, and style of learning for each
student.
The objective is to stay in the zone of proximal development
(Murray and Arroyo 2002), making the learning suggestions
not too easy but also not too hard. If the learning suggestions
are too easy, the student’s experience gradually becomes
boring. If they are too challenging, the student may become
increasingly frustrated. In a way, these programs are similar
to computer games, which are designed to keep the players
interested, occupied, and even addicted to the game, all of this
being critical for the game’s success.
Many big publishing companies invest millions of dollars
in adaptive learning programs that complement popular
textbooks. These applications are available either online or
on desktop computers (Daniel et al. 2015). The companies are
compelled to invest heavily in adaptive learning programs,
fearing competition and hoping for a high dividend. This
is a new business model for publishing companies because
highly profitable textbooks are not selling as much as they
used to and AL access is the new source of profits. Access to
adaptive learning programs cannot be stolen as easily as free
PDFs of the textbook. More importantly, AL applications can
eventually replace textbooks as the main educational resource
of the future.
How do the students react? Two divergent
scenarios.
For a few semesters, all of my anatomy and physiology
students systematically used online AL resources provided by
the publisher. Every semester my classroom observations and
student comments were essentially the same. The feedback
seemed consistently inconsistent and ambiguous. Then
patterns started to appear.
As the class begins using AL resources, students usually
respond in surprisingly different ways. Some love the
experience and others hate it. The two common scenarios that
suggest possible reasons for this wide range of reactions are
described below.
Motivated, persistent students use the AL programs a lot,
reading and answering many questions, reviewing all the
suggested materials and thinking hard about the topic. In
other words, they do what the program expects them to do.
In the process, they supply the program with abundant data
of good quality. The program uses the data to learn about the
student, adapting to the student’s individual learning needs.
This makes the learning experience even more exciting and
rewarding for the student, further motivating the learner to
work hard.
continued on next page
 Volume 22, Issue 3 December 2018
 Can Modern AI replace teachers? Not so fast! Artificial Intelligence and Adaptive Learning: Personalized Education in the AI age
Now let us consider another group of students who are
easily distracted and bored. They are not very persistent and
motivated in learning. These students tend to skip practicing
altogether. Even if they start answering the questions, they
may stop early. Overall, they generate much less data than
motivated, persistent students. With scarce data, the AI
program cannot help these students very much, which makes
it even more difficult for students to complete the assignment.
Another problem with easily distracted students is that the
quality of the data is usually not very high. If these students
are forced to reluctantly answer all the questions, they may
start clicking on the answers at random, without reading or
thinking about them at all. This behavior is easy to anticipate
since the students are trying to complete the assignment as
fast as possible instead of trying to learn as much as possible.
These students may quickly become bored and frustrated.
However, randomly clicking on answers is not what the
program expects students to do. The program becomes
confused. It gives the student new questions that make as
little sense to the learner as the old ones did. The frustration
grows. What was supposed to be an easy way out – clicking
through the questions, as useless as that is – becomes
an increasingly difficult path. This may account for some
persistent and very personal student responses I had in my
anatomy and physiology classes, for example, “I hate this
program so much!”
In this case, the low quantity and poor quality of student data
makes AI “blind.” It becomes as dumb and ineffective as it can
be (“garbage in, garbage out”). This also explains, to some
extent, why there are always a few students who benefit from
AI adaptive learning much more than others.
The paradox is clear: adaptive learning may help the most
the ones who need this help least. It also may help the least
the ones who need this help most. As a result, instead of
decreasing the gap between the high achieving and low
achieving students, these AI tools can actually increase the
gap.
Interestingly, it is the same AL game for all, but for some it
is exciting and even addictive and for others it is boring and
tedious. It very much depends on how you play it from the
start. The harder you try, the better it becomes. The more
you skip and cheat, the harder it is to play it. The program can
adapt to the learner, but the learner needs to adapt as well, in
order to get the greatest benefits.
The AL experiences of my students were program-specific,
but the AL achievement gap paradox is not. This hypothesis is
rooted in the basic understanding of the modern data-driven
AI, machine learning. The user provides the data. Good data
means effective AI. Poor data means ineffective AI. It cannot
be otherwise.
251 • HAPS Educator Journal of the Human Anatomy and Physiology Society
Unfortunately, there are no rigorous studies of the AL and
achievement gap in the current literature. Hopefully, close
attention to this problem will help improve performance of the
AI educational systems. Students who need more assistance
should get more, not less, from AL programs.
Can modern AI replace teachers?
The AL programs discussed in this article cannot replace an
experienced educator. These programs can be useful learning
tools, especially for some students. They can help teachers by
providing additional student training that frees classroom time
for higher level learning activities. However, adaptive learning
programs can also be easily confused by corrupt or scarce
student data. Adaptive learning programs do not understand
the wide range of the all-important pedagogical contexts.
They do not know how to deal with students who habitually
avoid learning or are afraid of learning and have never had a
positive learning experience. The student models created by
adaptive learning programs may be too simplistic and they are
often unrealistic.
Teachers, on the other hand, have always been engaged in
their own version of adaptive learning. We adapt our teaching
to the particular class and to individual students, as much as
we can, based on our experience, intuition, testing results and
the available “bag of tricks,” which has the limited selection of
educational resources and techniques we have at our disposal.
The feedback we use for creating our own “student models”
includes very important non-verbal clues in the classroom,
such as facial expression, posture, movement, immediate
reactions, and interactions in groups. We build personal
working relationships with many students and respond in class
and online with good humor and grace.
Can AI emulate all of this in the future? The next generation
adaptive learning programs are on the way (Lane et al. 2016,
Luckin et al. 2016). They will be much more sophisticated in
their understanding of pedagogical context and in the range
of potential adaptations to the learning needs of the student.
It is possible that future adaptive learning pedagogical
programs will change the way we learn.
About the author
Vasiliy Kolchenko is a professor of biology at New York City
College of Technology, The City University of New York.
Vasiliy teaches anatomy and physiology, pathophysiology,
and bioinformatics. He is developing a new introductory
course in artificial intelligence and neuroscience. His research
includes biosensor development and graphic representation
in science education. He also writes and performs music.
This is his Teaching Science song: https://www.youtube.com/
watch?v=CpeI5wHvKE4
continued on next page
 Volume 22, Issue 3 December 2018
 Can Modern AI replace teachers? Not so fast! Artificial Intelligence and Adaptive Learning: Personalized Education in the AI age

Literature cited Hawkins J, Blakeslee S (2007) On Intelligence: How a New

Understanding of the Brain will Lead to the Creation of
Andriessen J and Sandberg J (1999) Where is education
Truly Intelligent Machines. Macmillan.
heading and how about AI. International Journal of
Artificial Intelligence in Education 10.2: 130-150. Lane C, Gover S, Roschelle J (2016) AI Applications in
Education. Menlo Park, CA, USA: The Center for Innovative
Beck J, Mia S, Haugsjaa E (1996) Applications of AI in
Research in CyberLearning.
education. Crossroads 3.1: 11-15.
Luckin R, Holmes W, Griffiths M, Forcier L (2016) Intelligence
Bostrom N (2014) Superintelligence: Paths, Dangers, Strategies.
unleashed: An argument for AI in education. London:
Oxford University Press.
Pearson Education.
Brusilovsky P, Peylo C (2003) Adaptive and intelligent web-
Murray T, Arroyo I (2002) Toward measuring and maintaining
based educational systems. International Journal of
the zone of proximal development in adaptive
Artificial Intelligence in Education (IJAIED), 13: 159-172.
instructional systems. International Conference on
Brusilovsky P, Millán E (2007) User models for adaptive Intelligent Tutoring Systems (pp. 749-458). Berlin,
hypermedia and adaptive educational systems. In: The Heidelberg: Springer.
Adaptive Web (pp. 3-35). Berlin, Heidelberg: Springer.
Nilsson NJ (2009) The Quest for Artificial Intelligence.
Daniel J, Esteban VC, Mercè GC (2015) The future of MOOCs: Cambridge University Press.
Adaptive learning or business model? International Journal
Russell SJ, Norvig P (2010) Artificial Intelligence: A Modern
of Educational Technology in Higher Education 12.1: 64-73.
Approach (Third Edition). Pearson Education Limited.
Domingos P (2015) The Master Algorithm: How the Quest for
Sleeman D, Brown JS (1982) Intelligent Tutoring Systems.
the Ultimate Learning Machine will Remake Our World.
London: Academic Press.
Basic Books.

Back to TOC

252 • HAPS Educator Journal of the Human Anatomy and Physiology Society Volume 22, Issue 3 December 2018
 INTEGRATED HARDWARE, SOFTWARE & CURRICULUM
FOR LIFE SCIENCE EDUCATION

The choice of top labs worldwide

Use for introductory lessons, graduate porgrams, or advanced research

The ultimate teaching solution for...
Physiology Exercise Physiology & Biomechanics
Biology Psychophysiology & Neurophysiology
Biomedical Engineering Pharmacology & Toxicology
Nursing Veterinary Programs
When you get the BSL...they’ll get science!
Incorporate clincial techniques & industry standard equipement to teach your
students real-world skills.
ECG Biofeedback Nerve Conduction
EEG Bioimpedance Pulmonary Function
EMG Cardiac Output Reaction Time
EOG BP & Heart Sounds Respiration
EDA (GSR) Metabolic Rate Temperature
Multi-level learning solutions let you control the material and method of each
experiment.
• Inquiry-based student lessons • PRO Lessons
• Instructor modified labs • Custom Lessons
• Advanced analysis tools • No programming required

REQUEST A DEMO TODAY!

TEL 1+ 805-685-0066 WWW.BIOPAC.COM
FAX 1+ 805-685-0067 INFO@BIOPAC.COM

View publication stats